llvmpipe: implement support for multiple viewports

Largely related to making sure the rasterizer can correctly
pick out the correct scissor box for the current viewport.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca<jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

draw: implement support for multiple viewports

This adds support for multiple viewports to the draw module.
Multiple viewports depend on the presence of geometry shaders
which can write the viewport index.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: José Fonseca<jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

gallium: Add support for multiple viewports

Gallium supported only a single viewport/scissor combination. This
commit changes the interface to allow us to add support for multiple
viewports/scissors.

Signed-off-by: Zack Rusin <zackr@vmware.com>
Reviewed-by: Marek Olšák <maraeo@gmail.com>
Reviewed-by: José Fonseca<jfonseca@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

mesa: Delete the ctx->Array._RestartIndex derived state.

It's incorrect and isn't used any longer.

v2: Actually flush vertices/flag _NEW_TRANSFORM on RestartIndex change.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Ignore fixed-index primitive restart in ArrayElement().

GL_PRIMITIVE_RESTART_FIXED_INDEX is only supposed to apply to
glDrawElements*.  This code is for legacy drawing paths and display
lists, so it shouldn't apply.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

st/mesa: Go back to using ctx->Array.RestartIndex, not _RestartIndex.

The derived _RestartIndex field is an attempt to support both
GL_PRIMITIVE_RESTART and GL_PRIMITIVE_RESTART_FIXED_INDEX (part of ES
3.0).  Gallium drivers don't appear to support ES 3.0 yet, so they don't
need to use it.  Plus, it's broken and going to go away soon.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Fix can_cut_index_handle_restart_index() for byte/short types.

Pre-Haswell hardware doesn't support an arbitrary restart index, and
instead compares the index buffer value against 0xFF for byte-size
buffers, 0xFFFF for short-size buffers, or 0xFFFFFFFF for unsigned
integer buffers.

OpenGL allows the restart index to be an arbitrary unsigned integer.
When comparing against byte/short types, the index buffer value should
be promoted to a full 32-bit integer before doing the comparison.  The
restart index is /not/ supposed to be masked to byte/short size.

This means that with certain restart indexes, the comparison should
always fail.  For example, a restart index of 0xF000FFFF should never
match any byte/short index buffer values due to the extra high bits.

We must not enable hardware primitive restart in such a case.  For now,
fall back to software primitive restart as it's the simplest fix.  In
the future, we could detect restart indexes that will never match and
skip both hardware and software primitive restart.

NOTE: This is a candidate for stable branches.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965: Use the correct restart index for fixed index mode on Haswell.

The code that updates the ctx->Array._RestartIndex derived state mashed
it to 0xFFFFFFFF when GL_PRIMITIVE_RESTART_FIXED_INDEX was enabled
regardless of the index buffer type.  It's supposed to be 0xFF for byte,
0xFFFF for short, or 0xFFFFFFFF for integer types.

The new _mesa_primitive_restart_index() helper gets this right.

The hardware appears to compare against the full 32-bit value some of
the time, causing primitive restart not to occur when it should.  The
fact that it works some of the time is rather frightening.

Fixes sporadic failures in the ES 3 instanced_arrays_primitive_restart
conformance test when run in combination with other tests.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

vbo: Use the new primitive restart index helper function.

This gets the correct restart index for unsigned byte/short types when
using GL_PRIMITIVE_RESTART_FIXED_INDEX.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

mesa: Add a helper function for determining the restart index.

The derived state approach currently used (_RestartIndex) doesn't work:
in the GL_PRIMITIVE_RESTART_FIXED_INDEX case, the restart index depends
on the index buffer's data type, and that isn't known until draw time.

The existing code also fails to obey the GL 4.3 rules which say that
FIXED_INDEX takes precedence over normal primitive restart.

This helper function correctly determines the restart index, and will
replace the derived state.

NOTE: This is a candidate for the 9.1 branch.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

vbo: Ignore PRIMITIVE_RESTART_FIXED_INDEX for glDrawArrays().

The derived _PrimitiveRestart enable flag combines the PrimitiveRestart
and PrimitiveRestartFixedIndex enable flags.  However, DrawArrays is not
supposed to do FixedIndex restart:

From the OpenGL 4.3 Core specification, section 10.3.5 (page 302):
"If PRIMITIVE_RESTART_FIXED_INDEX is enabled, primitive restart is not
 performed for array elements transferred by any drawing command not
 taking a type parameter, including all of the *Draw* commands other
 than *DrawElements*."

The OpenGL ES 3.0 specification agrees by omission:
"When DrawElements, DrawElementsInstanced, or DrawRangeElements
 transfers a set of generic attribute array elements to the GL..."

Notably, DrawArrays is not included in the list of draw calls that
take PRIMITIVE_RESTART_FIXED_INDEX into consideration.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

i965/vs: Fix implied_mrf_writes() for integer division pre-gen6.

Previously it would assertion fail in debug builds (though the correct
value was returned in a non-debug build).  Marking it as a candidate for
stable even though it has no current consumers in the stable branches, in
case one shows up in a later backport.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64727
NOTE: This is a candidate for stable branches.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Fix test for smearing enabled on an instruction.

We were expanding the live range too far, breaking register_coalesce_2()
and compute_to_mrf() on 16-wide shaders.  Turning it back on improves
GLB2.7 performance by 0.239355% +/- 0.0850649% (n=398). shader-db stats
are:

total instructions in shared programs: 1627211 -> 1609262 (-1.10%)
instructions in affected programs:     450351 -> 432402 (-3.99%)

While 33 new 16-wide shaders are gained, 70 are lost.  Despite that,
tropics (the app that lost the most 16-wide) shows a .41% +/- .16%
(n=7/8, first-run outlier removed) performance improvement on my HSW.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965/fs: Fix segfault in instruction scheduling with LINTERP using last GRF.

The scheduler didn't know about uniform-type accesses, and if a uniform
access was last in a 16-wide, we'd walk off the end of the array.  This
never happened, because we'd never coalesce out all the GRFs, due to a bug
to be fixed in the next commit.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

mesa: Fix test for optimistic coloring being necessary.

i965 and radeon use ra_set_node_reg() to force payload registers to
specific registers while exposing those registers to the allocator still.
We were treating those register nodes as unsuccessfully allocated in the
ra_simplify() step, leading to walking the registers again to do
optimistic coloring even if there was nothing left ot do.

Acked-by: Kenneth Graunke <kenneth@whitecape.org>

gallium: fix build on uclibc system

execinfo.h and debug_symbol_name_glibc() are pure GNU-isms and do not
build on uclibc systems.  A previous patch addressed this issue, but
there was an error.  This patch corrects that error.  See

  https://bugs.freedesktop.org/show_bug.cgi?id=51782
  https://bugs.gentoo.org/show_bug.cgi?id=469768

Signed-off-by: Anthony G. Basile <blueness@gentoo.org>
Signed-off-by: Brian Paul <brianp@vmware.com>

intel: Enable blit glCopyTexSubImage/glBlitFramebuffer with sRGB.

Since the introduction of default-to-SARGB8 window system framebuffers,
non-blorp hardware lost blit acceleration for these two paths between the
window system and ARGB8888 textures.  Since we shouldn't be doing any
conversion anyway, just compatibility-check the linear variants of the
formats.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61954
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Tested-by: Tobias Jakobi <tjakobi@math.uni-bielefeld.de>

radeonsi: Add ipo to LLVM_COMPONENTS

r600g needs it too, so add ipo in the common radeon_llvm_check().

radeonsi compiled and linked, but it failed at dynamic link time
with a missing symbol.

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

llvmpipe: get rid of tiled/linear layout remains

Eliminate the rest of the no longer needed layout logic.
(It is possible some code could be simplified a bit further still.)

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

intel: Remove dead intel_drawbuf_region().

Since the glBitmap() MRT change, it's unused.  There was basically no way
to responsibly use this function since MRT was introduced.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Fix format handling of blit glBitmap()

Any 32-bit format got ARGB8888 handling (including, say, GL_RG1616), and
anything else got 16-bit (including, say, GL_R8), which could potentially
hang the GPU by writing out of bounds.

NOTE: This is a candidate for the stable branches.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Fix MRT handling of glBitmap().

We'd only hit color buffer 0 even if multiple draw buffers were bound.

NOTE: This is a candidate for the stable branches.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Rebuild PBO blit glTexImage() on top of miptrees.

This will ensure that we have resolves if we ever extend this to
glTexSubImage(), and fixes missing image start offset handling.

The texture buffer alloc ended up getting moved up, because we want to
look at the format of the image's actual mt to see if we'll end up
blitting the right thing, in the case of packed depth/stencil uploads.

This is the last caller of intelEmitCopyBlit() on a miptree-wrapped BO.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Rebuild PBO blit glReadPixels() on top of miptrees.

The previous code was missing depth resolves, that had only been prevented
due to no blitting of Y tiling.  The pair of flip args in the new blit
function means that we can just drop the pack->Invert fallback.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Rework intel_miptree_create_for_region() to wrap a BO.

I needed to do this for the PBO blit cases to use intel_miptree_blit().
But this also actually partially fixes a bug in EGLImage handling: We
can't share regions across contexts, because regions have a refcount that
isn't protected by a mutex, and different contexts can be simulataneously
accessed from multiple threads.  Now we just need to get regions out of
__DRIImage.  There was also a missing use of image->offset in the EGLImage
renderbuffer storage code.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Make a temporary miptree for the blit path of miptree mapping.

In a bit of debug code, we no longer have the inter-slice x/y to print.
But I think the level/slice is more useful in this case for looking at
what's getting mapped, especially given that INTEL_DEBUG=blit will tell
you the other value.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Make a temporary miptree when doing blit uploads for glTexSubImage().

While this is a bit more CPU work, it also is less code to handle this
path, and fixes problems with 32k-pitch textures and missing resolves.

v2: Add error checking in new code.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Extend the force_y_tiling flag to allow forcing no tiling.

For a blit-uploaded temporary, it's faster on current hardware to memcpy
the data into a linear CPU mapping than to go through the GTT.

v2: Turn the not-fully-supported mask into 3 supported enum values.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com> (v1)
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> (v1)
Reviewed-by: Paul Berry <stereotype441@gmail.com> (v2)
Reviewed-by: Chad Versace <chad.versace@linux.intel.com> (v2)

intel: Add an assert for glCopyTexSubImage() being called on MSAA buffers.

This is just in case someone else trips over this due to our weird reuse
of this code in glBlitFramebuffer().

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

i965: Allow glCopyTexSubImage() on depth textures.

If the hw is pre-gen5 and can't blit depth, it'll cleanly error out.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

i965: Prefer blorp glBlitFramebuffer() to the glCopyTexSubImage-based blit.

I think we've measured no performance difference from this in the past,
except that the blorp code can do things like multisample resolves.
Prevents piglit regression in the next commit when a testcase started
trying to do a multisampled resolve through the old glCopyTexSubImage()
path.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

i965: Consistently do depth resolves before blitting.

We were protected for a long time by the fact that depth was Y tiled and
you couldn't blit Y.  Now that we can blit Y, we were failing to resolve
depth in glCopyPixels().

Note in the comment about swrast, that the swrast map path does resolves
appropriately already.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Make a wrapper for intelEmitCopyBlit using miptrees.

I had previously asserted that it was hard to write a useful, simpler
blit function, but I think this might be it.

This has the side effect of extending the 32k pitch check to a few more
places that were missing it.

v2: Update comment for being moved inside intel_miptree_blit().

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Rename intel_renderbuffer_tile_offsets.

This makes it more consistent with intel_miptree_get_tile_offsets().

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Reduce intel_renderbuffer_tile_offsets to a thin wrapper.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Acked-by: Paul Berry <stereotype441@gmail.com>

intel: Make intel_miptree_get_tile_offsets return a page offset.

Right now, the callers in i965 don't expect a nonzero page offset to
actually occur (since that's being handled elsewhere), but it seems
like a trap to leave it this way.

Reviewed-and-tested-by: Ian Romanick <ian.d.romanick@intel.com>
Acked-by: Paul Berry <stereotype441@gmail.com>

glsl: Fix MSVC build.

It appears that `sizeof(Class::member)` is either non-standard or
merely unsupported in MSVC.

So use `sizeof(instance->member)` instead, which is guaranteed to work
everywhere.

Also promote the assert to a static assert.

Trivial.

mesa: fix GLSL program objects with more than 16 samplers combined

The problem is the sampler units are allocated from the same pool for all
shader stages, so if a vertex shader uses 12 samplers (0..11), the fragment
shader samplers start at index 12, leaving only 4 sampler units
for the fragment shader. The main cause is probably the fact that samplers
(texture unit -> sampler unit mapping, etc.) are tracked globally
for an entire program object.

This commit adapts the GLSL linker and core Mesa such that the sampler units
are assigned to sampler uniforms for each shader stage separately
(if a sampler uniform is used in all shader stages, it may occupy a different
sampler unit in each, and vice versa, an i-th sampler unit may refer to
a different sampler uniform in each shader stage), and the sampler-specific
variables are moved from gl_shader_program to gl_shader.

This doesn't require any driver changes, and it fixes piglit/max-samplers
for gallium and classic swrast. It also works with any number of shader
stages.

v2: - converted tabs to spaces
    - added an assertion to _mesa_get_sampler_uniform_value

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

swrast: increase array size of TextureSample

to match the size of ctx->Texture.Unit, and it will also fix
piglit/max-samplers with the following commit.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

mesa: declare UniformBufferBindings as an array with a static size

Some Gallium drivers were crashing, because the array was not large enough.

v2: clamp the per-shader maximum in st/mesa, then sum them all up

NOTE: This is a candidate for the stable branches.

radeonsi: Enable GLSL 1.30

radeonsi: Handle TGSI TXQ opcode

radeonsi: Add support for TGSI TXF opcode

radeonsi: Use tgsi_util_get_texture_coord_dim()

radeonsi: Handle TGSI_SEMANTIC_CLIPDIST

radeonsi: Make border colour state handling safe for integer textures

radeonsi: Fix hardware state for dual source blending

Set up CB_SHADER_MASK register according to pixel shader exports, and enable
some minimal state for colour buffer 1 in case dual source blending is used.

r600g/sb: handle more cases for folding in gvn pass

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

st/vdpau: destroy handle table only when it's empty

Signed-off-by: Christian König <christian.koenig@amd.com>

st/vdpau: remove vlCreateHTAB from surface functions

Signed-off-by: Christian König <christian.koenig@amd.com>

st/vdpau: invalidate the handles on destruction

Fixes a problem with xbmc when switching channels.

Signed-off-by: Christian König <christian.koenig@amd.com>

r600g/sb: improve folding for SETcc

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: optimize CNDcc instructions

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: improve optimization of conditional instructions

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

ilo: enable multiple constant buffers

This effectively enables uniform buffer object support.

ilo: add support for indirect access of CONST in FS

Unlike other register files, CONST is read with a message and indirect access
is easier to implement.

ilo: add support for TBOs on GEN6

This hunk was missing in the last commit.

ilo: advertise supports for pure integer formats

For pure integer formats, no filtering nor blending is needed.

ilo: add support for texture buffer objects

Take care of sampler views that have buffers as the underlying resources.
Update caps related to TBOs.

tgsi: add buffer texture to tgsi_util_get_texture_coord_dim()

TGSI_TEXTURE_BUFFER is one-dimensional.  Assert that exec_tex() is never
called with TGSI_TEXTURE_BUFFER.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

r600g/sb: improve handling of KILL instructions

This patch improves handling of unconditional KILL instructions inside
the conditional blocks, uncovering more opportunities for if-conversion.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: fix peephole optimization for PRED_SETE

Fixes incorrect condition that prevented optimization for
PRED_SETE/PRED_SETE_INT.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: fix scheduling of PRED_SET instructions

PRED_SET instructions that update exec mask should be scheduled immediately
prior to the "if-then-else" block, because any instruction that is
inserted after alu clause with PRED_SET and before conditional block is
also conditionally executed by hw (exec mask is already updated at that
moment).

Propbably it's better to make PRED_SET a part of conditional
"if-then-else" block in the IR to handle this more cleanly,
but for now this temporary solution should prevent the problem.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: fix handling of preloaded inputs for compute shaders

For compute shaders we need to let the backend know that
GPRs 0 and 1 are preloaded with some compute-specific input
values, otherwise any use of these regs without previous
definition is considered as undefined value and usually
is simply replaced with 0.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

xlib: add null ctx check in glXDestroyContext()

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64934
NOTE: This is a candidate for the stable branches.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

st/glx: add null ctx check in glXDestroyContext()

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64934
NOTE: This is a candidate for the stable branches.

Reviewed-by: Jose Fonseca <jfonseca@vmware.com>

st/mesa: add switch cases for new IR enums to silence warnings

st/glx/xlib: assorted whitespace, comment fixes

r600g/sb: fix incorrect assert

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: relax some restrictions for FETCH instructions

This allows GVN rewrite pass to propagate non-const (register)
values to FETCH source operands, helping to eliminate unnecessary
copies in some cases.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: relax register allocation for compute shaders

We have to assume that all GPRs in compute shader can be indirectly
addressed because LLVM backend doesn't provide any indirect array info.
That's why for compute shaders GPR array is created that covers all used
GPRs (0..r600_bytecode::ngpr-1), but this seriously restricts register
allocation in sb.

This patch checks for actual use of indirect access in the shader and
if it's not used then GPR array is not created, so that regalloc is not
unnecessarily restricted.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: fix gpr array handling for compute shaders

Fixes segfault with bfgminer and R600_DEBUG=sbcl.

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/sb: fix buffer overflow in sb_ostream

Fixes segfault during bytecode dump with bfgminer kernel

Signed-off-by: Vadim Girlin <vadimgirlin@gmail.com>

r600g/compute: Use common transfer_{map,unmap} functions for global resources

Reviewed-by: Marek Olšák <maraeo@gmail.com>

r600g/compute: Use common transfer_{map,unmap} functions for kernel inputs

Reviewed-by: Marek Olšák <maraeo@gmail.com>

i965: Go back to using the kernel SOL reset feature.

It turns out the MI_LOAD_REGISTER_IMM approach doesn't work on Haswell,
and regressed essentially all the transform feedback Piglit tests.

This morally reverts eaa6fbe6d54dc99efac4ab8e800edef65ce8220d.  However,
the code is still simpler than it was.  On BeginTransformFeedback, we
simply flush the batch and set the SOL reset flag so that the next batch
will start with zeroed offsets.  There's still no software counting.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=64887
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Paul Berry <stereotype441@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

freedreno: scissor fix

Don't assume the state-tracker will set the scissor after the
framebuffer state is changed.

Signed-off-by: Rob Clark <robclark@freedesktop.org>

freedreno: implement pipe->resource_copy_region()

Signed-off-by: Rob Clark <robclark@freedesktop.org>

glsl linker: compare interface blocks during interstage linking

Verify that interface blocks match when linking separate shader
stages into a program.

Fixes piglit glsl-1.50 tests:
* linker/interface-blocks-vs-fs-member-count-mismatch.shader_test
* linker/interface-blocks-vs-fs-member-order-mismatch.shader_test

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>

glsl linker: compare interface blocks during intrastage linking

Verify that interface blocks match when combining compilation
units at the same stage. (For example, when merging all vertex
shaders.)

Fixes piglit glsl-1.50 test:
* linker/interface-blocks-multiple-vs-member-count-mismatch.shader_test

v5 (Ken): Rename to link_interface_blocks.cpp and drop the separate .h
file for consistency with other linker code.  Remove "ok" variable.
Fold cross_validate_interface_blocks into its caller.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>

glsl linker: support arrays of interface block instances

With this change we now support interface block arrays.
For example, cases like this:

out block_name {
    float f;
} block_instance[2];

This allows Mesa to pass the piglit glsl-1.50 test:
* execution/interface-blocks-complex-vs-fs.shader_test

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl link_varyings: link interface blocks using the block name

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl linker: remove interface block instance names

Convert interface blocks with instance names into flat
interface blocks without an instance name.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl ast_to_hir: support in/out for interface blocks

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl ast_to_hir: reject row/column_major for in/out interface blocks

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl ast_to_hir: move uniform block symbols to interface blocks namespace

Uniform/interface blocks are a separate namespace from types.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl_symbol_table: add interface block namespaces

For interface blocks, there are three separate namespaces for
uniform, input and output blocks.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl parser: allow in & out for interface block members

Previously uniform blocks allowed for the 'uniform' keyword
to be used with members of a uniform blocks. With interface
blocks 'in' can be used on 'in' interface block members and
'out' can be used on 'out' interface block members.

The basic_interface_block rule will verify that the same
qualifier type is used with the block and each member.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl ast_to_hir: reject interpolation qualifiers for uniform blocks

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl parser: handle interface block member qualifier

An interface block member may specify the type:
in {
    in vec4 in_var_with_qualifier;
};

When specified with the member, it must match the same
type as interface block type.

It can also omit the qualifier:
uniform {
    vec4 uniform_var_without_qualifier;
};

When the type is not specified with the member,
it will adopt the same type as the interface block.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl parser: on desktop GL require GLSL 150 for instance names

Interface blocks in GLSL 150 allow an instance name to be used.

v2:
 * use state->check_version

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl parser: reject VS+in & FS+out interface blocks

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: parse in/out types for interface blocks

Previously only 'uniform' was allowed for uniform blocks.

Now, in/out can be parsed, but it will only be allowed for
GLSL >= 150.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl parser: rename uniform block to interface block

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

glsl: rename ast_uniform_block to ast_interface_block

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>

i965: Enable guardband clipping on Gen4/5.

Enables guardband clipping when the viewport covers the entire render
target.

No piglit regressions on Ironlake.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Eric Anholt <eric@anholt.net>

ARB_fp: accept duplicate precision options

Relaxes the validation of

   OPTION ARB_precision_hint_{nicest,fastest};

to allow duplicate options. The spec says that both /nicest/ and
/fastest/ cannot be specified together, but could be interpreted
either way for respecification of the same option.

Other drivers (NVIDIA etc) accept this, and at least one Unity3D game
expects it to succeed (Kerbal Space Program).

V2: Add spec quote.

Signed-off-by: Chris Forbes <chrisf@ijw.co.nz>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>

ilo: Initialize need_flush in draw_vbo.

need_flush was uninitialized if hw3d->new_batch was true.

Fixes "Uninitialized scalar variable" defect reported by Coverity.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>

radeon: Initialize variables in radeon_llvm_context_init.

'type' was not fully initialized when calling lp_build_context_init.

Fixes "Uninitialized scalar variable" defect reported by Coverity.

NOTE: This is a candidate for the stable branches.

Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

intel: Count fragments in our blitter-based glBitmap() path.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=59440
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>